LoopGrid for CrewAI

Record CrewAI agent model completions, tool executions, policy decisions, and human reviews as cryptographically verifiable LoopGrid evidence. Connects to your LoopGrid Core service; the bridge observes CrewAI events without executing business actions.

Install

pip install loopgrid-crewai

The integration has passed native CrewAI and signed LoopGrid Core end-to-end tests for the versions documented in VALIDATION.md.

What this does

CrewAI publishes lifecycle events using BaseEventListener and the global event bus. This adapter observes:

• LLMCallCompletedEvent → model_completed with CrewAI call_id, model, agent/task provenance, response commitment
• ToolUsageFinishedEvent when uncached/unblocked → tool_requested, then tool_executed as paired, post-fact evidence
• policy and delegated authority → only explicit application facts supplied to start_decision() / record_policy()
• application-supplied human review → only explicit record_human_review() call
• real-world outcome → only explicit record_outcome() call after authoritative downstream observation

It does not execute tools, gate actions, infer success from a tool response, or infer financial/physical effects.

Tool event semantics. CrewAI's pre/post tool hooks can run for a blocked action, and ToolUsageStartedEvent alone cannot prove execution. The adapter deliberately records tool evidence when a completed, uncached CrewAI tool event is observed. Tool start and completion times remain native event facts, not invented ordering claims. A cached result is not treated as a new execution. Denied and modified hooks were validated in real CrewAI runtime tests on RC4. General guarantees across every CrewAI tool implementation are not claimed.

Develop or test from source

Python 3.10–3.13 (CrewAI currently excludes Python 3.14):

python -m venv ..\crewai-spike
..\crewai-spike\Scripts\Activate.ps1
python -m pip install -e ".[dev]"
python -m pip show crewai loopgrid
python -m pytest -v
python -m compileall -q src examples tests
python .\examples\native_runtime_smoke.py

Target versions: crewai==1.15.25, loopgrid==0.8.0, Core 0.8.1-design-partner.

Connect to a CrewAI Crew

from crewai import Agent, Crew, Task, Process
from loopgrid_crewai import LoopGridCrewAI

bridge = LoopGridCrewAI(
    base_url="http://127.0.0.1:8000", workspace_id="default",
    agent_id="billing-crew",
)
bridge.install()  # register CrewAI native BaseEventListener once

decision = bridge.start_decision(
    decision_type="customer_refund",
    agent={"id":"billing-crew", "version":"1"},
    authority={"acting_for":"Example Store", "scope":["refund:create"], "limit_usd":100},
    model={"provider":"openai", "name":"configured-model"},
    context={"prompt_version":"refund-v1"},
    proposed_action={"tool":"refund.create", "amount":25},
    policy={"policy_id":"refund-policy", "version":"1", "decision":"auto_allowed"},
    metadata={"sandbox":True, "real_money_moved":False},
)

# Create your actual CrewAI Crew using the application's agents/tasks/tools.
crew = Crew(agents=[...], tasks=[...], process=Process.sequential)
with bridge.decision_context(decision["decision_id"]):
    result = crew.kickoff()
bridge.flush()
bridge.assert_healthy()  # transport success does not imply evidence completeness

# ONLY after an authoritative downstream system actually reports the outcome:
bridge.record_outcome(decision["decision_id"], {"status":"succeeded", "external_reference":"actual-external-id"}, observer="billing-webhook")

Important: A model-generated statement or CrewOutput alone is NOT an authoritative downstream business outcome. The example above shows the API, not a self-contained runnable Crew. For a runnable real Crew without API credentials see examples/native_runtime_smoke.py.

Human review and guardrails

Use CrewAI's native pre-tool execution hooks to block or modify actions. Supply authenticated reviewer information from your own application; after approval, explicitly invoke:

bridge.record_human_review(decision_id, reviewer="reviewer-id", approved=True, reason="Approved by operator")

This drains earlier queued evidence before the explicit review. The tool event listener is observation-only and is not an approval system. Denied and modified tool hooks passed Windows native runtime checks. Reviewer authentication remains application-owned.

Privacy / failure handling

capture_content=False by default. Tool arguments and results plus LLM responses are stored as SHA-256 commitments. Explicit application-provided metadata, agent roles, tool names and policy values may be visible; do not put secrets there. Opting into capture_content=True exports raw content.

The listener queues events to one FIFO worker, never intentionally blocks execution for network I/O, and retains transport failures for assert_healthy() after flush(). An error-free Crew execution is not proof that evidence was recorded successfully.

Correlation and limitations

decision_context uses Python contextvars. It isolates nested and parallel task contexts when propagated. Detached threads without propagated context are ignored rather than misattributed. Each invocation has short-lived per-scope ordinal/deduplication state and a random local run ID; an application may reopen the same decision without colliding with prior run keys. State becomes collectible when the context and dispatched handlers finish. Duplicate windows are capped at 2,048 tool objects and 2,048 LLM call identities per active scope. Replayed events older than the window may not be suppressed. CrewAI tool-completion events do not expose a stable provider tool-call ID, and this integration does not claim cross-process exactly-once delivery or deduplication of a newly constructed replayed tool event.

Core only marks a decision evidence_complete when its applicable evidence is actually recorded. A verify.valid=true result means evidence integrity, not business truth or safety.

Executable examples

Run inside a disposable virtual environment installed with python -m pip install -e ".[dev]".

The reviewer used by E2E scripts is a deterministic fixture, not a production-authenticated reviewer. The application remains responsible for real authentication and authorization. Neither capture_content=False nor raw SHA-256 commitments make low-entropy input unrecoverable; avoid recording secrets in explicit metadata, roles, or tool names.

Listener lifetime: create one bridge per intended application scope, call shutdown() after scoped runs and flush() are finished. shutdown() cannot be reversed; use a new bridge if required. Active-scoped deduplication is bounded; redacted transport error history retains the most recent 256 entries. CrewAI owns a process-global event bus, so handler registrations may remain as inert weak-owner callbacks until process exit.

Important limitations: completion-derived tool evidence is an ordered post-fact pair, not proof of pre-execution interception; no cross-process exactly-once action identity is promised. Multi-agent/hierarchical, deeply concurrent, provider-backed, cached-result, and detached-thread behavior must be checked separately before broad compatibility claims. Successful verification proves record integrity, not correctness of business facts.

Release gates

See VALIDATION.md for the Windows RC4 native, Core, distribution, and clean-wheel test results. GitHub CI and any public PyPI deployment are separately verified at release time.

Apache-2.0.